feat(vision): vehicle stage, phase A — YOLOX-S (Apache-2.0 ONNX) beside the plate recognizer

Fills /analyze vehicle.body_type + confidence (car / motorcycle / bus / truck from COCO,
mapped to the shared vocabulary) for the Car Wash desk's category suggestion
(venue-modules.md §Vehicle category from vision). Advisory: the operator decides, a
confident downgrade is flagged, nothing is gated on it.

- vision_service/vehicle.py: pure numpy/cv2 letterbox (pad 114, raw BGR), stride-grid
  decode, class-agnostic NMS, one vehicle per frame (the box holding the plate's centre,
  else the largest); YoloxVehicleDetector on onnxruntime CPU, 2 intra-op threads.
- recognizer.py: WithVehicle composes the stage over any plate recognizer (stub included);
  a failing stage yields vehicle=null + a "vehicle: …" note in /health.detail — never
  costs the plate read. model_version reads "<plate>+yolox:yolox_s.onnx@640".
- settings: VISION_VEHICLE_MODEL_PATH (unset = off), _INPUT_SIZE (640), _MIN_CONFIDENCE
  (0.4, the detector's floor; the flag threshold is site config).
- Dockerfile bakes yolox_s.onnx (best-effort curl at build; no network → stage off) and
  sets the path; compose forwards it (empty = off); .env.example documents it.
- Measured on four real dev entry frames (DS-2CD1047G3H, 2560×1440): car at 0.83–0.88 in
  ~240–330 ms; empty lane with a person → none.
- tests/test_vehicle.py: decode/NMS/pick/letterbox on synthetic tensors, the composition,
  and a missing-model /health. Wiki: opencv-anpr-service, venue-modules, log.

Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU

apps/vision/.env.example

@@ -20,3 +20,12 @@ VISION_OCR_MODEL=cct-xs-v2-global-model
 # Confidence floor — a best plate below this is flagged low_confidence so the Node side
 # treats it as advisory and falls back to the ticket path. Keep in sync with the server.
 VISION_MIN_CONFIDENCE=0.5
+
+# Vehicle stage (phase A): a YOLOX ONNX graph (Apache-2.0) run on the same frame after the
+# plate read; fills /analyze `vehicle.body_type` (car/truck/bus/motorcycle) + confidence for
+# the Car Wash desk's category suggestion. Unset = off. The Docker image bakes the weights
+# at /app/models/yolox_s.onnx; locally: curl the release file into apps/vision/models/.
+#   https://github.com/Megvii-BaseDetection/YOLOX/releases/download/0.1.1rc0/yolox_s.onnx
+# VISION_VEHICLE_MODEL_PATH=models/yolox_s.onnx
+# VISION_VEHICLE_INPUT_SIZE=640
+# VISION_VEHICLE_MIN_CONFIDENCE=0.4

apps/vision/Dockerfile

@@ -14,7 +14,7 @@ ENV UV_LINK_MODE=copy \
 
 # System libs the recognizer stack needs (opencv/onnxruntime): GL + glib. Kept minimal.
 RUN apt-get update \
-    && apt-get install -y --no-install-recommends libgl1 libglib2.0-0 \
+    && apt-get install -y --no-install-recommends libgl1 libglib2.0-0 curl \
     && rm -rf /var/lib/apt/lists/*
 
 # ---- deps: resolve + install the venv from the lockfile (cache-friendly) ----
@@ -26,6 +26,13 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # ---- project source ----
 COPY vision_service/ ./vision_service/
 COPY README.md ./
+# Vehicle stage weights (phase A): YOLOX-S, Apache-2.0, ~36 MB, baked into the image so the
+# air-gapped appliance never fetches at runtime and no operator-writable path holds a model
+# (vision-service-hardening.md). Best-effort at build: without network the stage stays off.
+ARG YOLOX_URL=https://github.com/Megvii-BaseDetection/YOLOX/releases/download/0.1.1rc0/yolox_s.onnx
+RUN mkdir -p /app/models \
+    && (curl -fsSL -o /app/models/yolox_s.onnx "$YOLOX_URL" \
+        || (echo "[build] yolox weights not fetched (no network) — vehicle stage off" && rm -f /app/models/yolox_s.onnx))
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv sync --frozen --extra alpr
 
@@ -49,7 +56,8 @@ RUN uv run python -c "from fast_alpr import ALPR; ALPR()" \
 # Default to the stub recognizer (offline, no model load); override to fast_alpr in prod.
 ENV VISION_RECOGNIZER=stub \
     VISION_HOST=0.0.0.0 \
-    VISION_PORT=8089
+    VISION_PORT=8089 \
+    VISION_VEHICLE_MODEL_PATH=/app/models/yolox_s.onnx
 EXPOSE 8089
 HEALTHCHECK --interval=30s --timeout=5s --start-period=20s --retries=3 \
   CMD python -c "import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://localhost:8089/health').status==200 else 1)" || exit 1

apps/vision/tests/test_vehicle.py

@@ -0,0 +1,143 @@
+"""Vehicle stage (phase A) — pure post-processing on synthetic tensors, and the
+recognizer composition over the stub with a fake detector. No weights needed."""
+
+from __future__ import annotations
+
+import os
+
+import numpy as np
+from fastapi.testclient import TestClient
+
+from vision_service.schemas import BBox, VehicleResult
+from vision_service.vehicle import (
+    COCO_VEHICLE_CLASSES,
+    Detection,
+    decode,
+    letterbox,
+    nms,
+    pick_vehicle,
+    vehicles_from_output,
+)
+
+SIZE = 64  # tiny "model" input: grids 8x8 + 4x4 + 2x2 = 84 rows
+ROWS = (SIZE // 8) ** 2 + (SIZE // 16) ** 2 + (SIZE // 32) ** 2
+
+
+def raw_output(hits: list[tuple[int, int, int, float, float, float, float]]) -> np.ndarray:
+    """Build a YOLOX-style raw tensor [ROWS, 85] with the given (row, coco_class, _, obj,
+    cls_score, log_w, log_h) hits; everything else is background."""
+    raw = np.zeros((ROWS, 85), dtype=np.float32)
+    raw[:, 2:4] = -10.0  # exp → ~0 size for background rows
+    for row, cls, _, obj, score, lw, lh in hits:
+        raw[row, 0:2] = 0.5  # centre of its grid cell
+        raw[row, 2] = lw
+        raw[row, 3] = lh
+        raw[row, 4] = obj
+        raw[row, 5 + cls] = score
+    return raw
+
+
+def test_decode_maps_grid_offsets_and_log_sizes_to_pixels() -> None:
+    raw = raw_output([(0, 2, 0, 1.0, 1.0, np.log(2.0), np.log(3.0))])
+    dec = decode(raw, SIZE)
+    # Row 0 = stride-8 grid cell (0,0): centre (0.5+0)*8 = 4, size exp(log 2)*8 = 16 / 24.
+    assert dec[0, :4].tolist() == [4.0, 4.0, 16.0, 24.0]
+    # Last row = stride-32 cell (1,1): centre (0.5+1)*32 = 48.
+    raw2 = raw_output([(ROWS - 1, 7, 0, 1.0, 1.0, 0.0, 0.0)])
+    dec2 = decode(raw2, SIZE)
+    assert dec2[ROWS - 1, :4].tolist() == [48.0, 48.0, 32.0, 32.0]
+
+
+def test_vehicles_only_above_floor_mapped_to_vocabulary_and_scaled_back() -> None:
+    raw = raw_output(
+        [
+            (0, 2, 0, 0.9, 0.9, np.log(2.0), np.log(2.0)),  # car, score .81
+            (1, 0, 0, 0.99, 0.99, np.log(2.0), np.log(2.0)),  # person → ignored
+            (2, 7, 0, 0.5, 0.5, np.log(2.0), np.log(2.0)),  # truck, score .25 → below floor
+        ]
+    )
+    found = vehicles_from_output(raw, SIZE, scale=0.5, min_confidence=0.4)
+    assert [d.body_type for d in found] == ["car"]
+    assert round(found[0].confidence, 2) == 0.81
+    # Box 16px wide in the letterboxed input → 32px in the original (scale 0.5).
+    assert round(found[0].x2 - found[0].x1) == 32
+    assert set(COCO_VEHICLE_CLASSES.values()) == {"car", "motorcycle", "bus", "truck"}
+
+
+def test_nms_keeps_the_best_of_overlapping_boxes() -> None:
+    boxes = np.array([[0, 0, 10, 10], [1, 1, 11, 11], [50, 50, 60, 60]], dtype=np.float32)
+    scores = np.array([0.5, 0.9, 0.7], dtype=np.float32)
+    assert sorted(nms(boxes, scores, 0.45)) == [1, 2]
+
+
+def test_pick_prefers_the_box_holding_the_plate_else_the_largest() -> None:
+    near = Detection("car", 0.9, 0, 0, 100, 100)
+    far = Detection("truck", 0.8, 200, 200, 400, 400)  # larger
+    inside = Detection("car", 0.7, 10, 10, 60, 60)  # tighter box also holding the plate
+    assert pick_vehicle([near, far], None) is far
+    assert pick_vehicle([near, far], BBox(x1=20, y1=20, x2=30, y2=30)) is near
+    assert pick_vehicle([near, far, inside], BBox(x1=20, y1=20, x2=30, y2=30)) is inside
+    assert pick_vehicle([near, far], BBox(x1=900, y1=900, x2=910, y2=910)) is far  # plate outside every box
+    assert pick_vehicle([], None) is None
+
+
+def test_letterbox_keeps_aspect_and_pads_with_114() -> None:
+    frame = np.zeros((30, 60, 3), dtype=np.uint8)
+    tensor, scale = letterbox(frame, 64)
+    assert tensor.shape == (1, 3, 64, 64) and tensor.dtype == np.float32
+    assert abs(scale - 64 / 60) < 1e-9
+    assert tensor[0, 0, 63, 63] == 114.0  # padding
+    assert tensor[0, 0, 0, 0] == 0.0  # image
+
+
+class FakeDetector:
+    model_version = "fake-vehicle"
+
+    def __init__(self, result: VehicleResult | None) -> None:
+        self.result = result
+        self.calls: list[BBox | None] = []
+
+    def detect(self, image_bytes: bytes, plate: BBox | None) -> VehicleResult | None:
+        self.calls.append(plate)
+        return self.result
+
+
+def test_composition_fills_vehicle_over_the_stub_and_survives_a_failing_stage() -> None:
+    from vision_service.recognizer import StubRecognizer, WithVehicle
+    from vision_service.settings import Settings
+
+    det = FakeDetector(VehicleResult(body_type="truck", confidence=0.77))
+    rec = WithVehicle(StubRecognizer(Settings()), det)
+    res = rec.analyze(b"jpeg-bytes")
+    assert res.plate is None
+    assert res.vehicle == VehicleResult(body_type="truck", confidence=0.77)
+    assert res.model_version == "stub-0+fake-vehicle"
+    assert det.calls == [None]
+
+    class Boom:
+        model_version = "boom"
+
+        def detect(self, image_bytes: bytes, plate: BBox | None) -> VehicleResult | None:
+            raise RuntimeError("no model")
+
+    rec2 = WithVehicle(StubRecognizer(Settings()), Boom())
+    res2 = rec2.analyze(b"jpeg-bytes")
+    assert res2.vehicle is None
+    assert rec2.ready is True
+    assert "vehicle: RuntimeError: no model" in (rec2.error or "")
+
+
+def test_app_reports_a_missing_model_file_and_keeps_serving() -> None:
+    from vision_service.app import app
+
+    os.environ["VISION_VEHICLE_MODEL_PATH"] = "/nonexistent/yolox.onnx"
+    try:
+        with TestClient(app) as client:
+            health = client.get("/health").json()
+            assert health["ready"] is True  # the plate stage (stub) is fine
+            assert "vehicle:" in (health["detail"] or "")
+            res = client.post("/analyze", content=b"x", headers={"content-type": "application/octet-stream"})
+            assert res.status_code == 200
+            assert res.json()["vehicle"] is None
+    finally:
+        os.environ.pop("VISION_VEHICLE_MODEL_PATH", None)

apps/vision/vision_service/recognizer.py

@@ -19,6 +19,7 @@ from typing import Protocol
 
 from .schemas import AnalyzeResponse, BBox, PlateResult
 from .settings import Settings
+from .vehicle import VehicleDetector, YoloxVehicleDetector
 
 
 class Recognizer(Protocol):
@@ -165,10 +166,59 @@ class FastAlprRecognizer:
         )
 
 
+class WithVehicle:
+    """Composition: any plate recognizer + the vehicle stage. Runs the plate stage first
+    (its box picks WHICH vehicle), then fills `vehicle`. A failing vehicle stage is
+    logged into `error` and yields null — it must never cost the plate read."""
+
+    def __init__(self, inner: Recognizer, detector: VehicleDetector) -> None:
+        self._inner = inner
+        self._detector = detector
+        self.vehicle_error: str | None = None
+
+    @property
+    def model_version(self) -> str:
+        return f"{self._inner.model_version}+{self._detector.model_version}"
+
+    @property
+    def ready(self) -> bool:
+        return bool(self._inner.ready)
+
+    @property
+    def error(self) -> str | None:
+        inner = getattr(self._inner, "error", None)
+        det = getattr(self._detector, "error", None) or self.vehicle_error
+        parts = [p for p in (inner, f"vehicle: {det}" if det else None) if p]
+        return "; ".join(parts) if parts else None
+
+    def analyze(self, image_bytes: bytes) -> AnalyzeResponse:
+        started = time.perf_counter()
+        res = self._inner.analyze(image_bytes)
+        try:
+            vehicle = self._detector.detect(image_bytes, res.plate.bbox if res.plate else None)
+        except Exception as exc:  # noqa: BLE001 - advisory stage, never fatal
+            self.vehicle_error = f"{type(exc).__name__}: {exc}"
+            vehicle = None
+        took_ms = (time.perf_counter() - started) * 1000.0
+        return res.model_copy(
+            update={"vehicle": vehicle, "model_version": self.model_version, "took_ms": took_ms}
+        )
+
+
 def build_recognizer(settings: Settings) -> Recognizer:
     """Factory: pick the recognizer from settings. Falls back to the stub if the real
-    one can't load, so the service always comes up (with ready=False surfaced)."""
+    one can't load, so the service always comes up (with ready=False surfaced). The
+    vehicle stage wraps whichever recognizer runs when a model path is configured."""
+    rec: Recognizer
     if settings.recognizer == "fast_alpr":
         rec = FastAlprRecognizer(settings)
+    else:
+        rec = StubRecognizer(settings)
+    if settings.vehicle_model_path:
+        detector = YoloxVehicleDetector(
+            settings.vehicle_model_path,
+            input_size=settings.vehicle_input_size,
+            min_confidence=settings.vehicle_min_confidence,
+        )
+        return WithVehicle(rec, detector)
     return rec
-    return StubRecognizer(settings)

apps/vision/vision_service/settings.py

@@ -32,6 +32,16 @@ class Settings(BaseSettings):
     # Node side can fall back to the ticket path rather than trust it.
     min_confidence: float = 0.5
 
+    # Vehicle stage (phase A — venue-modules.md §Vehicle category from vision): a YOLOX
+    # ONNX graph (Apache-2.0) run beside the plate recognizer. Unset = stage off (the
+    # response's `vehicle` stays null). Bake the file into the image (models/), never a
+    # path an operator can write (vision-service-hardening.md).
+    vehicle_model_path: str | None = None
+    vehicle_input_size: int = 640
+    # Detection score floor for a vehicle box to count at all (the Node side applies the
+    # site's own, stricter threshold before it FLAGS anything).
+    vehicle_min_confidence: float = 0.4
+
 
 def get_settings() -> Settings:
     return Settings()

apps/vision/vision_service/vehicle.py

@@ -0,0 +1,247 @@
+"""Vehicle stage: a COCO object detector beside the plate recognizer (Job 2, phase A).
+
+Answers "what KIND of vehicle is in this entry frame?" for the Car Wash desk's category
+suggestion (wiki/decisions/venue-modules.md §Vehicle category from vision). ADVISORY by
+design: the Node server records it next to the plate, the desk pre-selects the site
+category it maps to, the operator decides, a confident downgrade is flagged. Nothing is
+ever gated on it, so a wrong or missing detection costs nothing but a suggestion.
+
+Model: YOLOX (Megvii, Apache-2.0) as an ONNX graph on the ONNX Runtime the plate stage
+already uses — the licence rule that keeps Ultralytics (AGPL) out. COCO's vehicle classes
+are car / motorcycle / bus / truck: enough to tell a van or a truck from a car, NOT enough
+for SUV vs sedan — that is phase B (a body-type classifier on the pilot's own frames).
+The detector's vehicle box is also the crop phase B will classify.
+
+Pure numpy/cv2 pre/post-processing, no torch: letterbox to the model's square input
+(pad 114, no normalisation — YOLOX's exported graphs take raw 0–255 BGR), decode the
+stride grids, class-agnostic NMS, map COCO ids to the shared vocabulary, pick ONE
+vehicle: the one whose box holds the plate (when a plate was read), else the largest.
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Protocol
+
+from .schemas import BBox, VehicleResult
+
+# COCO-80 class index → the shared VEHICLE_CLASSES vocabulary (packages/shared).
+COCO_VEHICLE_CLASSES: dict[int, str] = {2: "car", 3: "motorcycle", 5: "bus", 7: "truck"}
+
+# YOLOX feature strides; grids are input/stride per level (8400 anchors at 640).
+_STRIDES = (8, 16, 32)
+
+
+@dataclass(frozen=True)
+class Detection:
+    body_type: str
+    confidence: float
+    x1: float
+    y1: float
+    x2: float
+    y2: float
+
+    @property
+    def area(self) -> float:
+        return max(0.0, self.x2 - self.x1) * max(0.0, self.y2 - self.y1)
+
+    def contains(self, x: float, y: float) -> bool:
+        return self.x1 <= x <= self.x2 and self.y1 <= y <= self.y2
+
+
+class VehicleDetector(Protocol):
+    """What the recognizer composition needs: frame bytes (+ the plate box) → a class."""
+
+    @property
+    def model_version(self) -> str: ...
+
+    def detect(self, image_bytes: bytes, plate: BBox | None) -> VehicleResult | None: ...
+
+
+# ----------------------------------------------------------------------------------
+# Pre/post-processing (pure functions — unit-tested on synthetic tensors)
+# ----------------------------------------------------------------------------------
+
+
+def letterbox(frame: Any, size: int) -> tuple[Any, float]:
+    """Resize keeping aspect, pad bottom/right with 114 to size×size. Returns the CHW
+    float32 tensor (batch dim added) and the scale to map boxes back."""
+    import cv2
+    import numpy as np
+
+    h, w = frame.shape[:2]
+    r = min(size / h, size / w)
+    nh, nw = int(round(h * r)), int(round(w * r))
+    resized = cv2.resize(frame, (nw, nh), interpolation=cv2.INTER_LINEAR)
+    padded = np.full((size, size, 3), 114, dtype=np.uint8)
+    padded[:nh, :nw] = resized
+    tensor = padded.transpose(2, 0, 1)[None].astype(np.float32)
+    return np.ascontiguousarray(tensor), r
+
+
+def decode(raw: Any, size: int) -> Any:
+    """YOLOX raw output [N, 5+classes] (batch squeezed) → same shape with xywh decoded
+    into pixel units of the letterboxed input. Rows are ordered stride 8, 16, 32."""
+    import numpy as np
+
+    out = raw.astype(np.float32).copy()
+    grids = []
+    strides = []
+    for s in _STRIDES:
+        n = size // s
+        ys, xs = np.meshgrid(np.arange(n), np.arange(n), indexing="ij")
+        grids.append(np.stack((xs, ys), axis=-1).reshape(-1, 2))
+        strides.append(np.full((n * n, 1), s, dtype=np.float32))
+    grid = np.concatenate(grids, axis=0).astype(np.float32)
+    stride = np.concatenate(strides, axis=0)
+    if out.shape[0] != grid.shape[0]:
+        raise ValueError(f"unexpected output rows {out.shape[0]} for input {size} (want {grid.shape[0]})")
+    out[:, :2] = (out[:, :2] + grid) * stride
+    out[:, 2:4] = np.exp(out[:, 2:4]) * stride
+    return out
+
+
+def nms(boxes: Any, scores: Any, iou_threshold: float) -> list[int]:
+    """Greedy class-agnostic non-max suppression over xyxy boxes; returns kept indices."""
+    import numpy as np
+
+    if len(boxes) == 0:
+        return []
+    order = scores.argsort()[::-1]
+    x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
+    areas = np.clip(x2 - x1, 0, None) * np.clip(y2 - y1, 0, None)
+    keep: list[int] = []
+    while order.size > 0:
+        i = int(order[0])
+        keep.append(i)
+        if order.size == 1:
+            break
+        rest = order[1:]
+        xx1 = np.maximum(x1[i], x1[rest])
+        yy1 = np.maximum(y1[i], y1[rest])
+        xx2 = np.minimum(x2[i], x2[rest])
+        yy2 = np.minimum(y2[i], y2[rest])
+        inter = np.clip(xx2 - xx1, 0, None) * np.clip(yy2 - yy1, 0, None)
+        iou = inter / (areas[i] + areas[rest] - inter + 1e-9)
+        order = rest[iou <= iou_threshold]
+    return keep
+
+
+def vehicles_from_output(
+    raw: Any, size: int, scale: float, min_confidence: float, iou_threshold: float = 0.45
+) -> list[Detection]:
+    """Full post-processing: decode → vehicle classes only → confidence floor → NMS →
+    boxes in ORIGINAL frame pixels."""
+    import numpy as np
+
+    dec = decode(raw, size)
+    cls_scores = dec[:, 5:]
+    cls_idx = cls_scores.argmax(axis=1)
+    score = dec[:, 4] * cls_scores[np.arange(len(dec)), cls_idx]
+    wanted = np.isin(cls_idx, list(COCO_VEHICLE_CLASSES)) & (score >= min_confidence)
+    if not wanted.any():
+        return []
+    d = dec[wanted]
+    s = score[wanted]
+    c = cls_idx[wanted]
+    boxes = np.stack(
+        (d[:, 0] - d[:, 2] / 2, d[:, 1] - d[:, 3] / 2, d[:, 0] + d[:, 2] / 2, d[:, 1] + d[:, 3] / 2), axis=1
+    )
+    keep = nms(boxes, s, iou_threshold)
+    out: list[Detection] = []
+    for i in keep:
+        b = boxes[i] / scale
+        out.append(
+            Detection(
+                body_type=COCO_VEHICLE_CLASSES[int(c[i])],
+                confidence=float(s[i]),
+                x1=float(b[0]),
+                y1=float(b[1]),
+                x2=float(b[2]),
+                y2=float(b[3]),
+            )
+        )
+    return out
+
+
+def pick_vehicle(detections: list[Detection], plate: BBox | None) -> Detection | None:
+    """ONE vehicle per frame: the box holding the plate's centre (the car that was read —
+    a lane frame can show the car behind too), else the largest box (nearest the camera)."""
+    if not detections:
+        return None
+    if plate is not None:
+        cx = (plate.x1 + plate.x2) / 2
+        cy = (plate.y1 + plate.y2) / 2
+        holders = [d for d in detections if d.contains(cx, cy)]
+        if holders:
+            return min(holders, key=lambda d: d.area)  # the tightest box around the plate
+    return max(detections, key=lambda d: d.area)
+
+
+# ----------------------------------------------------------------------------------
+# The ONNX Runtime detector
+# ----------------------------------------------------------------------------------
+
+
+class YoloxVehicleDetector:
+    """YOLOX ONNX on onnxruntime (CPU). Loads once; a load failure is surfaced through
+    `error` and the stage simply yields no vehicle (never breaks the plate path)."""
+
+    def __init__(self, model_path: str, input_size: int = 640, min_confidence: float = 0.4) -> None:
+        self._path = Path(model_path)
+        self._size = input_size
+        self._min_confidence = min_confidence
+        self._session = None
+        self._input_name = "images"
+        self._error: str | None = None
+        try:
+            import onnxruntime as ort
+
+            opts = ort.SessionOptions()
+            opts.intra_op_num_threads = 2  # one frame per entry; leave cores to the lane
+            self._session = ort.InferenceSession(
+                str(self._path), sess_options=opts, providers=["CPUExecutionProvider"]
+            )
+            self._input_name = self._session.get_inputs()[0].name
+        except Exception as exc:  # noqa: BLE001 - not-ready, never fatal
+            self._error = f"{type(exc).__name__}: {exc}"
+
+    @property
+    def model_version(self) -> str:
+        return f"yolox:{self._path.name}@{self._size}"
+
+    @property
+    def ready(self) -> bool:
+        return self._session is not None
+
+    @property
+    def error(self) -> str | None:
+        return self._error
+
+    def detect(self, image_bytes: bytes, plate: BBox | None) -> VehicleResult | None:
+        if self._session is None:
+            return None
+        import cv2
+        import numpy as np
+
+        frame = cv2.imdecode(np.frombuffer(image_bytes, dtype=np.uint8), cv2.IMREAD_COLOR)
+        if frame is None:
+            return None
+        tensor, scale = letterbox(frame, self._size)
+        raw = self._session.run(None, {self._input_name: tensor})[0][0]
+        found = vehicles_from_output(raw, self._size, scale, self._min_confidence)
+        best = pick_vehicle(found, plate)
+        if best is None:
+            return None
+        return VehicleResult(body_type=best.body_type, confidence=round(best.confidence, 4))
+
+
+def time_detect(
+    detector: VehicleDetector, image_bytes: bytes, plate: BBox | None
+) -> tuple[VehicleResult | None, float]:
+    """detect() with wall time in ms (for logs/benchmarks)."""
+    started = time.perf_counter()
+    result = detector.detect(image_bytes, plate)
+    return result, (time.perf_counter() - started) * 1000.0

docker-compose.yml

@@ -55,6 +55,9 @@ services:
     environment:
       # Engine: stub (no models) by default; prod override sets fast_alpr.
       VISION_RECOGNIZER: ${VISION_RECOGNIZER:-stub}
+      # Vehicle stage (Car Wash category suggestion): the image bakes YOLOX-S at this path.
+      # Set the var to an EMPTY value in the stack env to switch the stage off.
+      VISION_VEHICLE_MODEL_PATH: ${VISION_VEHICLE_MODEL_PATH-/app/models/yolox_s.onnx}
     networks:
       - parking
 

wiki/decisions/venue-modules.md

@@ -188,11 +188,13 @@ onto the site's own categories ("car, sedan, hatchback → Vetura").
   and the read is at or above the threshold → one `anomaly` (`carwash.categoryDowngrade`, both
   categories, both prices, operator, snapshotId) and `downgrade_event_id` on the order. Equal,
   upgrade, unsure or unmapped reads flag nothing. The order is always created.
-- **Model — NOT built.** No bundled recognizer produces `body_type` yet, so today the desk shows
-  nothing and nothing is flagged. Phase A = a COCO detector on Apache-2.0 ONNX weights (car /
-  truck / bus / motorcycle, plus the vehicle crop); Phase B = the body-type classifier trained on
-  the pilot's own frames — every wash order is a labelled frame (entry snapshot + the category a
-  person chose), so the dataset builds itself on park-2. Reports (discrepancies per operator per
+- **Model — phase A built (same day).** YOLOX-S (Apache-2.0 ONNX) as a vehicle stage beside the
+  plate recognizer: car / truck / bus / motorcycle + the vehicle crop, ~250 ms per entry frame on
+  CPU, weights baked into the vision image. Details and measurements on [[opencv-anpr-service]]
+  §Vehicle body type. Phase B = the body-type classifier trained on the pilot's own frames —
+  every wash order is a labelled frame (entry snapshot + the category a person chose), so the
+  dataset builds itself on park-2. Until then a Vetura/SUV list sees every car as Vetura and no
+  downgrade fires; van/truck/bus/motorcycle do separate. Reports (discrepancies per operator per
   shift) wait for the first real reads.
 
 ## Car Wash — the pilot module (settled 2026-09-05)

wiki/entities/opencv-anpr-service.md

@@ -249,11 +249,36 @@ service's `/health` each tick and shows a **"Vision" chip** in the booth footer
 - Per-camera **opt-in** — ✅ **built**: `config.anpr === true` enables ANPR on a camera (set via the
   SetupWizard checkbox); ANPR then runs on that camera's entry/exit snapshot.
 
-## Vehicle body type (advisory) — contract only, 2026-09-06
+## Vehicle body type (advisory) — the vehicle stage, phase A (2026-09-06)
 
-`/analyze` may now populate `vehicle.body_type` + `vehicle.confidence` from the shared vocabulary
+`/analyze` populates `vehicle.body_type` + `vehicle.confidence` from the shared vocabulary
 (car, sedan, hatchback, suv, minivan, pickup, van, truck, bus, motorcycle). Node records it beside
 the plate and the Car Wash desk pre-selects the category the site maps it to; the operator
-decides, a confident downgrade is flagged, nothing is gated on it. No bundled recognizer emits
-it yet — see [[venue-modules]] §Vehicle category from vision for the model plan (COCO detector
-first, body-type classifier on own frames second).
+decides, a confident downgrade is flagged, nothing is gated on it ([[venue-modules]] §Vehicle
+category from vision).
+
+**Phase A = YOLOX-S (Megvii, Apache-2.0) as ONNX** on the same ONNX Runtime the plate stage uses
+— the licence rule that keeps Ultralytics (AGPL) out. `vision_service/vehicle.py`: pure numpy/cv2
+letterbox (pad 114, raw 0–255 BGR — YOLOX's exported graphs are not normalised), stride-grid
+decode, class-agnostic NMS, COCO `car/motorcycle/bus/truck` → the vocabulary, and ONE vehicle per
+frame: the box holding the plate's centre when a plate was read (the car that was read, not the
+one behind), else the largest box. `WithVehicle` in `recognizer.py` wraps whichever plate
+recognizer runs (stub included, so the stage is testable without fast-alpr); a failing stage
+yields `vehicle: null` and a `vehicle: …` note in `/health.detail` — it never costs the plate
+read. Composed `model_version` reads `<plate>+yolox:yolox_s.onnx@640`.
+
+- **Config:** `VISION_VEHICLE_MODEL_PATH` (unset = stage off), `VISION_VEHICLE_INPUT_SIZE` (640),
+  `VISION_VEHICLE_MIN_CONFIDENCE` (0.4 — the detector's floor; the SITE threshold that decides a
+  flag lives in Setup → Car wash). The Docker image bakes the weights at `/app/models/yolox_s.onnx`
+  (best-effort curl at build; no network → stage off) and sets the path, so the air-gapped
+  appliance never fetches at runtime and no operator-writable path holds a model
+  ([[vision-service-hardening]]). Compose forwards the var; set it EMPTY in the stack env to
+  switch the stage off. Locally: curl the release file into `apps/vision/models/` (gitignored).
+- **Measured on dev (2026-09-06), four real 2560×1440 entry frames from the DS-2CD1047G3H:** three
+  with a car → `car` at 0.83–0.88, ~240–330 ms each on the dev CPU with 2 intra-op threads; the
+  empty-lane frame with a person at the camera → no vehicle (correct: a person is not a class we
+  keep). One frame per entry, so the cost is invisible to the lane.
+- **What it cannot do:** SUV vs sedan — COCO has one `car`. For a Vetura/SUV price list every car
+  maps to Vetura and no downgrade fires; vans, trucks, buses and motorcycles do separate. Phase B
+  (a body-type classifier on the pilot's own frames — every wash order is a labelled frame) is
+  what closes that gap; the detector's box is the crop it will classify.

wiki/log.md

@@ -3091,3 +3091,12 @@ category + threshold; the desk pre-selects the mapped category and shows the rea
 a confident, pricier-mapped read with a cheaper choice signs `anomaly carwash.categoryDowngrade`
 (both categories/prices, operator, snapshot) — never blocks. No recognizer emits body_type yet.
 Tests in carwash.test.ts. Updated [[venue-modules]] (As built), [[opencv-anpr-service]].
+
+## [2026-09-06] ingest | Vision vehicle stage, phase A: YOLOX-S beside the plate recognizer
+`vision_service/vehicle.py` (YOLOX ONNX on onnxruntime: letterbox, grid decode, NMS, COCO
+car/motorcycle/bus/truck → vocabulary, one vehicle per frame — the box holding the plate, else the
+largest) + `WithVehicle` composition over any plate recognizer; `VISION_VEHICLE_MODEL_PATH` (unset =
+off), input size, detector floor; Dockerfile bakes yolox_s.onnx (best-effort curl) and sets the
+path; compose forwards it (empty = off). Measured on four real dev entry frames: car at 0.83–0.88,
+~240–330 ms, empty lane → none. Tests: tests/test_vehicle.py (pure post-processing + composition +
+missing-model health). Updated [[opencv-anpr-service]], [[venue-modules]].